PatchScout tackles web research by deciding when to move on

A researcher identifying college autism support programs needs more than a convincing answer with a handful of examples. The goal is a reusable collection of evidence from many institutions, including programs that are difficult to distinguish from ordinary disability services. Relevant pages may be scattered across websites that do not link to one another.

Michael West and Eduard Dragut, of Temple University, call this problem thematic web data collection. Their preprint presents PatchScout, a multi-agent framework intended to gather relevant documents while balancing depth within sources against breadth across them.

Searching as information foraging

The framework draws on information foraging theory, which describes information seeking as movement between promising patches. In PatchScout, a patch consists of the pages returned by a search query and the nearby pages reachable through their links.

Staying in a productive patch can uncover useful material efficiently. Staying too long can consume the collection budget without adding many new sources. Following links also creates a rapidly expanding pool of candidates, making exhaustive exploration impractical.

PatchScout divides the work among three language-model agents. The Scout generates search queries and adapts them using previous results. The Forager ranks links using signals such as anchor text and surrounding content. The Critic judges whether visited pages meet the collection theme. All three use DeepSeek-V3.1 in the reported experiments.

The central control is a switching rule. Some variants leave a patch when its observed yield falls below a threshold. Others use “information scent”: estimates of how likely unexplored links are to lead to relevant pages. Reference thresholds can be fixed or updated from earlier patches.

What the experiments found

The researchers tested four themes: U.S. four-year universities, Fortune 500 employment pages, accredited U.S. law schools, and college autism support programs. Reference lists enabled them to estimate coverage of known entities, accounting for organizations that use multiple domains.

PatchScout variants and the traversal-based crawler ACHE received budgets of 10,000 page visits per run. The study also tested the open-source DeepResearcher agent and commercial Claude and Gemini deep-research systems. Those agents ran until they produced reports rather than under the same page budget, an important distinction when interpreting the comparisons.

The strongest PatchScout variant for universities reached 20.3% coverage of the reference set, compared with 3.6% for ACHE and 1.8% for the best-performing deep-research baseline. For Fortune 500 employment, its best coverage was 24.6%, versus 15.0% for the strongest deep-research baseline.

For autism support programs, the best PatchScout policy recovered 39.1% of the known set, compared with 17.3% for the strongest research-agent baselines. The authors also report manually verifying 102 additional program entities absent from their incomplete, partly outdated reference collection. These were newly identified for that collection, not necessarily newly established programs.

Law schools provided a counterexample to a blanket superiority claim. Gemini reached 93.4% coverage and Claude 91.9%, exceeding PatchScout’s best result of 86.8%. The authors attribute this advantage to an authoritative accreditation directory that largely enumerates the target population: locating a comprehensive list can make sustained exploration unnecessary.

More pages do not mean broader coverage

One important finding is the gap between collecting relevant pages and discovering distinct sources. On universities, a productivity-driven PatchScout variant classified 86.2% of visited pages as relevant but covered only 4.4% of the reference set. The variant with the broadest coverage had a lower relevant-page yield of 77.7%.

In other words, a collector can look highly productive while repeatedly exploring a narrow part of the web. Forward-looking link scores generally helped breadth, but no switching policy won across all themes. A fixed productivity threshold performed best for the particularly sparse autism-program task.

Costs and uncertainty remain

Full PatchScout runs consumed substantially more language-model tokens than DeepResearcher. A separate analysis stopped PatchScout at DeepResearcher’s topic-specific token allowance; the best PatchScout variant then matched or exceeded its coverage on every topic. That does not establish equal-cost superiority over the commercial systems, whose usage reporting limited comparable analysis.

The evaluation also has uncertainty. Each configuration was tested in a single run, without error bars or significance tests, and changing webpages and search rankings limit exact reproducibility.

Relevance judgments are another qualification. A graded reassessment found that 25.8% of sampled pages originally labeled relevant were either irrelevant or merely topically related without supporting evidence. The authors therefore describe absolute harvest rates as optimistic at the margins.

The study supports a focused conclusion: deciding when to leave a web neighborhood can matter as much as choosing which link to follow. It does not show that any tested system has solved comprehensive collection.